This thesis presents a Discrete Element Method (DEM) to simulate irregular shaped particles by a non-convex polyhedron representation. By using novel GPU techniques and an efficient HPC implementation the presented method shows a level of throughput not previously attained with polyhedron particle representations in the open literature. Further, via such a representation the exact volumetric overlaps of the particles are resolved and, as a result, the method is robust and numerically stable with respect to geometric changes. The efficient and well-behaved method allows for significant progress in the study of granular materials, where previously mainly the inadequate particle representation of spherical or clumped spherical particles have been used.
The exact volumetric overlaps are resolved by a simplex representation which allows for the use of non-convex particles without any decomposition, aiding both performance and the ease of use of the method. Further, care is given to attain efficient scaling of the method with respect to particle resolution. Such a property enable for studies on higher resolution particles than previously shown in related work, and is result of efficient filtering of polyhedron triangles in the narrow contact phase. In addition, other novel techniques, such as a GPU BVH implementation for the broad phase contact detection, also aids the performance and the flexibility of the proposed and implemented method.
The method is shown to be convergent with respect to particle resolution, both for individual particle collisions and also for laboratory scale particle systems. The HPC implementation is proven to be highly efficient, where, for instance, a one second simulation of one million non-convex particles is simulated within an hour on a single GPU. By the effective filtering of triangles in the narrow contact phase, near linear scaling can be achieved with regards to particle resolution.